Journal of International Financial Markets,

Institutions and Money 12 (2002) 183– 200

www.elsevier.com/locate/econbase

On measuring volatility and the GARCH

forecasting performance
Emilio Barucci a,*, Roberto Renò b
a
Dipartimento di Statistica e Matematica Applicata all’Economia, Uni6ersità di Pisa,
6ia Cosimo Ridolfi 10, 56124 Pisa, Italy
b
Scuola Normale Superiore, Piazza dei Ca6alieri 7, 56121 Pisa, Italy
Received 16 December 2000; accepted 20 June 2001

Abstract

We apply a new algorithm based on Fourier analysis to compute the volatility of a

diffusion process. By using simulations of the continuous-time GARCH model, we show that
our method performs well in computing integrated volatility. We show that linear interpola-
tion of high frequency observations induces a downward bias in estimating integrated
volatility. By measuring ex post volatility with our method, we find that the forecasting
performance of the GARCH model is improved with respect to what is established when
classical methods are employed. These results are confirmed by the analysis of exchange rate
high frequency time series. © 2002 Elsevier Science B.V. All rights reserved.

Keywords: Volatility; High frequency data; GARCH models

JEL classification: C22; C53; F31

1. Introduction

Volatility estimation and forecasting is an outstanding topic in the financial

literature. Indeed, it plays a crucial role in many different fields, e.g. risk manage-
ment, time series forecasting and contingent claim pricing.
In the last 20 years, starting out from empirical investigations showing that
volatility in financial time series is highly persistent with clustering phenomena,

* Corresponding author. Tel.: + 39-050-945-235; fax: + 39-050-945-375.

E-mail addresses: ebarucci@ec.unipi.it (E. Barucci), reno@cibs.sns.it (R. Renò).

1042-4431/02/$ - see front matter © 2002 Elsevier Science B.V. All rights reserved.
PII: S 1 0 4 2 - 4 4 3 1 ( 0 2 ) 0 0 0 0 2 - 1
184 E. Barucci, R. Renò / Int. Fin. Markets, Inst. and Money 12 (2002) 183–200

many models have been proposed to describe volatility evolution. The literature is
now quite large with several specifications of auto-regressive models. Empirical
analyses have shown a high degree of inter-temporal volatility persistence, but in
many papers it has also been observed that forecasting with GARCH models can
be extremely unsatisfactory when the daily volatility is measured ex post by the
squared (daily) return, see Andersen and Bollerslev (1998c), Andersen et al. (1998),
Figlewski (1997), Pagan and Schwert (1990). In Andersen and Bollerslev (1998c) it
is shown that the forecasting performance of a GARCH(1,1) model is improved
when the daily volatility (integrated volatility) is measured by means of the
cumulative squared intraday returns. Monte Carlo experiments of processes, whose
parameters have been estimated on exchange rate time series (DM–$ and Yen –$),
show that the noise of the high frequency volatility estimator is smaller than that
of the daily squared return. On this topic see also Andersen et al. (1999b, 2000),
Barndorff-Nielsen and Shephard (2000a,b).
In this paper we address volatility estimation and forecasting in a GARCH
setting with high frequency data by applying a new algorithm, which has been
developed in Malliavin and Mancino (2000), to compute the volatility of a diffusion
process. This method is based on Fourier analysis techniques (hereafter Fourier
method). The volatility of a diffusion process is defined as the limit of its quadratic
variation. This definition motivates standard volatility estimation methods based on
a differentiation procedure: the quadratic variation of a process with a given
frequency (day, week, month) is taken as a volatility estimate. Estimating volatility
by this method presents some drawbacks when high frequency data are used.
Tick-by-tick data are not equally spaced, in the above cited papers an equally
spaced time series for intraday returns is constructed by linearly interpolating
logarithmic midpoints of bid-ask adjacent quotes or by taking the last quote before
a given reference time (henceforth called imputation method). This procedure
induces some distortions in the analysis, e.g. it may be at the origin of returns
autocorrelation, and it reduces the number of observations. The method proposed
in this paper avoids these problems; it is based on integration of the time series, and
it employs all the (irregularly spaced) observations. We assume the price to be
piecewise constant, i.e. the price is constant between two subsequent observations.
Volatility computation by using all the data with the Fourier method should be
more precise. We illustrate this fact through Monte Carlo simulations of a
continuous-time GARCH(1,1) model with the parameters estimated in Andersen
and Bollerslev (1998c). We show that the variance of our integrated volatility
estimator is smaller than that of the cumulative squared intraday returns. The
precision of the cumulative squared intraday returns in measuring volatility de-
pends on the procedure employed to build an equally spaced time series. Linear
interpolation causes a downward bias, which increases with sampling frequency.
The imputation method is immune from these drawbacks. Linear interpolation
seems to be at the origin of a downward bias, and this conjecture is confirmed by
implementing the Fourier method with linearly interpolated observations instead of
assuming the price to be piecewise constant. In this case, a strong downward bias
arises.
 E. Barucci, R. Renò / Int. Fin. Markets, Inst. and Money 12 (2002) 183–200 185

Through Monte Carlo simulations, we show that, by measuring integrated

volatility according to the Fourier method, the forecasting performance of the
GARCH(1,1) model is better than that obtained by computing volatility according
to the cumulative squared intraday returns.
These results are confirmed when the method is applied to compute volatility of
exchange rate high frequency time series. We apply the Fourier method to the
evaluation of the forecasting performance of the daily GARCH model and of the
intraday GARCH model, as in Andersen et al. (1998). For both the time series
considered, the GARCH model forecasts are evaluated to be better if the Fourier
method is employed as a volatility estimate instead of the cumulative squared
intraday returns.
Our paper is organized as follows. In Section 2 we present the algorithm to
compute volatility. In Section 3 we address the algorithm implementation. In
Section 4 we compute volatility when the asset price process is described by a
GARCH(1,1) model, and by using Monte Carlo experiments we compare it to the
estimate obtained with the sum of squared intraday returns. In Section 5 we
evaluate through simulations the GARCH(1,1) model volatility forecasting perfor-
mance by measuring volatility with the new method. In Section 6 we evaluate the
daily GARCH model performance in forecasting daily volatility of exchange rate
high frequency time series. In Section 7 we consider the GARCH model forecasting
performance when it is extended to intraday returns. Section 8 concludes.

2. Measuring volatility via Fourier analysis

The algorithm developed in Malliavin and Mancino (2000) allows us to compute

the diffusion process volatility by Fourier analysis. In what follows we present the
method, for a detailed analysis we refer the reader to the paper.
Given a process p(t), we only require its quadratic variation to be bounded.
Among the processes satisfying this assumption, we have the diffusion process:
dp(t)=|(t)dW(t) + v(t)dt, (1)
where | and v are time dependent random functions and W(t) is a Brownian
motion.
The instantaneous volatility at time t of any process with well defined quadratic
variation can be defined as follows:
1
S2(t)œlim E[(p(t +  ) − p(t))2 Ft ], (2)
 ¡0 

where E[· Ft ] denotes the conditional expectation operator with respect to the
|-field Ft generated by the full observation of the process until time t; for model
(Eq. (1)) we have S2(t) = | 2(t). We normalize the time window [0, T] in which the
time series is recorded to [0, 2y]. In Malliavin and Mancino (2000) Theorem 1.2 it
is shown that the Fourier coefficients of | 2 can be computed by means of the
Fourier coefficients of dp, then classical results of Fourier theory allows us to
reconstruct | 2(t)Öt  [0, 2y]. The Fourier coefficients of dp are
 &
186 E. Barucci, R. Renò / Int. Fin. Markets, Inst. and Money 12 (2002) 183–200
2y
1
a0(dp)= dp(t)
2y
&
0
2y
1
ak (dp)= cos(kt)dp(t)
y
&
0
2y
1
bk (dp)= sin(kt)dp(t); k ] 1. (3)
y 0

Following Malliavin and Mancino (2000), we obtain the Fourier coefficients of

| 2 through the formulae:
y n
1
a0(| 2)= lim % [a 2s (dp) +b 2s (dp)] (4)
n“ n +1 − n0 s = n 0 2
n
2y
ak (| 2)= lim % as (dp)as + k (dp) (5)
n“ n +1 − n0 s = n 0
n
2y
bk (| 2)= lim % as (dp)bs + k (dp). (6)
n“ n +1 − n0 s = n 0
Note that we are left with the choice of omitting the first n0 coefficients, since
they could be too sensitive to the drift term in Eq. (1)1. By the classical Fourier–Fé-
jer inversion formula, we can reconstruct | 2(t):

| 2(t)= lim % 1 −
n
  k
[a (| 2)cos(kt) +bk (| 2)sin(kt)]. (7)
n“ k=0 n k
The generalization to multivariate volatility is straightforward, see Malliavin and
Mancino (2000). If p(t) is observed in continuous time, this method would allow us
to reconstruct the instantaneous volatility at any time inside the interval [0, 2y].

3. Implementation

Given a time series of N observations (ti, p(ti)), i= 1, …, N, we will compress the

data in the interval [0, 2y] and compute the integrals in Eq. (3) through integration
by parts:

ak (dp)=
1 & 2y
cos(kt)dp(t) =
p(2y) −p(0) k
−
& 2y
sin(kt)p(t)dt. (8)
y 0 y y 0

To implement the method and in particular the integration, we need an assump-

tion on how the data are connected. Our choice is to set p(t) equal to p(ti ) in the
interval [ti, ti + 1] (piecewise constant). Then, the integral in Eq. (8) in the interval [ti,
ti + 1] becomes

1
The fact that the lowest frequency coefficients could be affected by a trend component has been
argued by Malliavin and Mancino (2000). In our analysis, we will use n0 =1, since exchange rate high
frequency time series do not display systematic trends.
 & &
E. Barucci, R. Renò / Int. Fin. Markets, Inst. and Money 12 (2002) 183–200 187
ti + 1 ti + 1
k k 1
sin(kt)p(t)dt =p(ti ) sin(kt)dt = p(ti ) [cos(kti )− cos(kti + 1)].
y ti y ti y
(9)
The smallest wavelength that can be evaluated in Eq. (3) to avoid aliasing effects
is twice the smallest distance between two consecutive prices; in the case of equally
spaced data, it will correspond to the frequency N/2 (Nyquist frequency). This
frequency is the largest that can be evaluated. If the Fourier coefficients of | 2 Eqs.
(4)–(6) have to be evaluated for 05 k 5 J, then the largest frequency n in Eq. (7)
is
N
n = − J, (10)
2
because of the terms as + k, bs + k in Eqs. (5) and (6). In the integration by parts
formula Eq. (8), the constant term p(2y)− p(0)/2y appears in each coefficient; this
term can be set at zero by adding in Eq. (1) the drift term − [p(2y)–p(0)/2y]dt.
This change of variables will not affect the volatility estimate.
In Barucci et al. (2000) we tested this method on equally spaced data. Monte
Carlo experiments simulating a diffusion process with constant volatility showed
that the method allows to estimate volatility in a univariate setting and cross-
volatilities in a multivariate setting. The precision of the estimate is similar to that
of classical methods. When applied to the daily time series of the Dow Jones
Industrial and Dow Jones Transportation Index, the Fourier method replicates the
volatility estimates obtained by the classical method.

4. Volatility computation

Let p(t)= log S(t), where S(t) is a generic asset price. Following a large
literature, we model the asset price through the continuous-time GARCH model:
dp(t)= |(t)dW1(t)
d| 2(t)=q[ −| 2(t)]dt + 2uq| 2(t)dW2(t), (11)
where q, , u are constants and W1, W2 are two independent Brownian motions.
Provided |(t)dW(t) is a continuous martingale, our method allows for jumps in
the process |(t). Jumps inserted directly in the differential equation driving the
price evolution are not allowed.
Given a time window [0, 1] (a day, week, month), we wish to compute the
integrated volatility of the process, i.e. 10 | 2(t)dt. An unbiased estimator of this
quantity is provided by [p(1) − p(0)]2. However, this estimator is very noisy. In
Andersen and Bollerslev (1998c) for exchange rates and in Martens (2000) for a
stock index, it is shown that an estimator with smaller noise is provided by the sum
of squared intraday returns: N 2
i = 2[p(i/N) − p(i− 1/N)] . Using as evaluation crite-
rion the difference 0 | (t)dt −i = 2[p(i/N) −p(i −1/N)]2, the authors find better
1 2 N

results on simulated time series with the highest frequency considered, correspond-
ing to 5-minute returns.
188 E. Barucci, R. Renò / Int. Fin. Markets, Inst. and Money 12 (2002) 183–200

The method proposed in Malliavin and Mancino (2000) gives us an estimator of

integrated volatility. In fact, integrating | 2(t) between 0 and 2y we get
& 2y
| 2(t)dt =2ya0(| 2), (12)
0

where a0(| 2) is given by Eq. (4). The computation of a0(| 2) provides an estimate of
the integrated volatility using all the observations.
To illustrate the validity of the Fourier approach, we simulate the diffusion
process Eq. (11) by an Euler discretization scheme, see for example Kloeden and
Platen (1992). We use the parameters q =0.035, = 0.636, u= 0.296 estimated in
Andersen and Bollerslev (1998c) on the daily return time series of the Deutsch
Mark –US Dollar exchange rate. Existence of the exact discretization of the process
Eq. (11) is guaranteed by Drost and Werker (1996) in a weak sense. On this topic,
see also the comprehensive results on temporal aggregation in Meddahi and
Renault (2000).
Taking a day as a reference measure, we simulate 24 hours of trading with
dt= 1/86 400 when discretizing Eq. (11), which corresponds to an observation every
second. In order to simulate high frequency unevenly sampled observations, we
select a subset of [1, 86 400] by extracting the time differences from an exponential
distribution with the mean equal to ~= 14 s. This choice is motivated by the fact
that the empirical distribution of ti −ti − 1 can be approximated with an exponential
shape and 14 seconds is the mean duration between quotes in the DM–$ exchange
rate time series. As a result, we will have a data set (tk, p(tk ), k=1, …, N) with tk
unevenly sampled, and |(t) recorded every second, so that the generated value of
the integrated volatility can be computed. Then we compute the integrated volatility
according to three estimators: the squared daily return, the cumulative 5-minute
squared intraday returns with linear interpolation of adjacent observations as
implemented in Muller et al. (1990), and the Fourier estimator Eq. (12). Results are
illustrated in Fig. 1, where the distribution of the normalized difference between the
integrated volatility and its estimate is shown. As expected, the squared daily return
is a very noisy estimator. As argued in Andersen and Bollerslev (1998c), when
estimating the volatility with the cumulative squared intraday returns, we get a
smaller variance. However, measuring volatility according to the Fourier method
we have a further reduction of the variance, as well as of the measurement bias.
Increasing the sampling frequency of a financial time series we encounter
problems related to microstructure effects. In our setting (simulated time series)
such effects are not present; as a consequence, by increasing the return sampling
frequency from 5 minutes to, say, 1 minute the cumulative squared intraday return
estimator should perform better. This is definitely not the case. Fig. 1 shows that
increasing the frequency from 5 minutes to 1 minute the variance of the cumulative
squared intraday return estimator is reduced but a downward bias comes in.
Considering diffusion processes belonging to the SR-SARV(1) class, in Barucci and
Renò (2000) we show that, as the sampling frequency is increased, the cumulative
squared intraday return estimator with linearly interpolated returns is more and
more downward biased.
 E. Barucci, R. Renò / Int. Fin. Markets, Inst. and Money 12 (2002) 183–200 189

This downward bias does not depend on the estimation method. If we implement
the Fourier method by computing the integrals in Eq. (3) interpolating linearly the
prices in the interval [ti, ti + 1], instead of assuming the price to be constant, we get
again a downward biased estimator; on the simulation sample of Fig. 1, we get a
downward bias in the mean of 0.426. Linear interpolation induces spurious
autocorrelation; in some sense, a straight line is the ‘minimum variance’ path
between two observations.
The downward bias effect of linear interpolation was conjectured in Corsi et al.
(2001). In that paper the authors suggest to use an imputation algorithm which
coincides with our piecewise constant assumption. As a matter of fact, taking the
last observation before time t as p(t) is equivalent to assume p(t)= p(ti ) in the
interval [ti, ti + 1]. With this imputation scheme, the cumulative squared intraday
return estimator is unbiased, as shown in Fig. 2 where the cumulative squared
intraday returns with the adoption of the imputation scheme is computed for a
sampling frequency of 2 minutes, 1 minute and 14 seconds. As suggested by the
theory, increasing the sampling frequency the variance of this estimator decreases.
In the limit, it converges from above to the variance of the Fourier estimator.
The Fourier estimator is characterized by the smallest variance for any sampling
frequency, regardless of the adoption of the imputation or of the linear interpola-

Fig. 1. Distribution of 10 | 2(t)dt− |ˆ 2/ 10 | 2(t)dt where |ˆ 2 is obtained with four different estimators of
the integrated volatility: (a) |ˆ 2 = [p(1)− p(0)]2; (b) |ˆ 2 = 288 2
i = 2[p(i/288) −p(i− 1/288)] (5-minute estima-
tor); (c) |ˆ 2 = 1440
i=2 [p(i/1440)−p(i− 1/1440)] 2
(1-minute estimator); (d) |
ˆ 2
=2ya0 (| 2
) (Fourier estima-
tor). In (b) and (c) returns are linearly interpolated. For each distribution, we indicate its mean and its
standard deviation (Std). The distributions are computed with 10 000 ‘daily’ replications.
190 E. Barucci, R. Renò / Int. Fin. Markets, Inst. and Money 12 (2002) 183–200

Fig. 2. Distribution of 10 | 2(t)dt− |ˆ 2/ 10 | 2(t)dt where |ˆ 2 is obtained with four different estimators of
the integrated volatility: (a) 2-minute cumulative squared intraday returns; (b) 1-minute cumulative
squared intraday returns; (c) 14-second cumulative squared intraday returns; (d) Fourier estimator. In
(a), (b) (c) returns are obtained with an imputation scheme. For each distribution, we indicate its mean
and its standard deviation (Std). The distributions are computed with 10 000 ‘daily’ replications.

tion scheme for the cumulative squared intraday return estimator. In what follows,
when volatility is computed according to the cumulative squared intraday returns,
we will aggregate the data through linear interpolation, as it is done in large part
of literature.

5. Volatility forecasting evaluation

Following Drost and Nijman (1993), Drost and Werker (1996) the GARCH
continuous time diffusion Eq. (11) can be discretized, obtaining the weak
GARCH(1,1) process:

n 2t + 1 =„+ h · r 2t +i · n 2t , (13)

where rt =p(t)−p(t − 1) and n 2t is the best linear predictor of r 2t expressed as a

linear combination of lagged squared returns. In Drost and Werker (1996) the exact
relation between „, h, i and q, , u is provided, so that one can calibrate Eq. (13)
on a given time series to obtain the coefficients of the continuous time diffusion Eq.
(11).
 E. Barucci, R. Renò / Int. Fin. Markets, Inst. and Money 12 (2002) 183–200 191

In this setting, n 2t + 1 provides us with an unbiased forecast of 10 | 2(t+~)d~.

While there is a strong support in favor of a high persistence in the volatility
dynamics, the 1 day ahead forecasting performance of the above model has been
evaluated to be very poor in the literature. However, as outlined in Andersen and
Bollerslev (1998c), the reason for this apparently disappointing result hinges on the
fact that the volatility measure which was used to evaluate ex post the forecasting
performance was the squared daily return. These authors point out that a careful
measure of the realized integrated volatility is needed in order to correctly evaluate
the model forecasting performance.
Using simulated time series, we employ the Fourier method to measure the
realized integrated volatility, and we evaluate the GARCH(1,1) model forecasting
performance according to it. The GARCH model performance with respect to this
estimator is compared with that associated with the cumulative squared intraday
returns with linearly interpolated observations. Following Andersen and Bollerslev
(1998c), the forecasting performance of the GARCH model can be evaluated
according to the R 2 of the linear regression:

|ˆ 2t =a+ b·n 2t +  t (14)

which is given by

R 2 =[corr(|ˆ 2t ,n 2t )]2, (15)

where |ˆ 2 is the ex post integrated volatility estimator.

We use the simulation setting described in Section 4. For the Y–$ time series we
set the mean duration equal to 52. The forecasting model is given by (Eq. (13)) with
the parameters as in Table 3 for m =1, corresponding to those employed in the
above section according to Drost and Werker (1996).
We point out that the R 2 obtained in Eq. (15) must be compared with the R 2
obtained when the ex post integrated volatility measure is perfectly known; its value
is given by

R 2 = [corr(
& t
| 2(s)ds, n 2t )]2, (16)
t−1

a value that can be computed in our simulation setting.

Table 1
R 2 obtained on simulated data with different estimators of integrated volatility

Estimator R 2 (DM–$) R 2 (Y–$)

(p(1)−p(0))2 0.062 0.092

288
i = 2(p(i/288)−p(i−1/288))
2
0.476 0.488
Fourier 0.489 0.501
R2 0.491 0.505

The values are computed through 50 000 ‘daily’ replications.

192 E. Barucci, R. Renò / Int. Fin. Markets, Inst. and Money 12 (2002) 183–200

Fig. 3. 2ya0(| 2) computed according to Eq. (4) as a function of n for the DM – $ exchange rate (1
October 1992). The dashes line indicates the cut we perform to compute integrated volatility.

Results are shown in Table 1: the Fourier estimator gives an R 2, which is very
close to R 2 . Employing the Fourier estimator, the GARCH forecasting perfor-
mance is better than that obtained by measuring integrated volatility through the
sum of squared intraday returns.

6. Exchange rate time series analysis

The data set at hand consists of the 1 year (from 1 October 1992 to 30 September
1993) collection of tick-by-tick quotes (bid and ask) of the Deutsch Mark–US
Dollar exchange rate and of Japanese Yen–US Dollar exchange rate, with time
stamps rounded to the nearest even second. The data set was collected and
delivered to us by Olsen & Associates. This data set has been extensively studied,
e.g. see Andersen and Bollerslev (1997, 1998b,c), Andersen et al. (1999a, 1998),
Muller et al. (1990), Zumbach (2000).
We define the price to be the mid-price between bid and ask quotes. The trading
day is chosen to begin and to end at 21:00 GMT. We excluded weekends and
trading days with less than 1000 quotes for the DM–$ and less than 320 quotes for
the Y –$. Moreover we applied the filter described in Dacorogna et al., (1993) which
removes roughly 0.36% of the quotes. We end up with 1 466 944 quotes for the
DM –$ and 567 758 quotes for Y– $, distributed over 258 trading days.
When applying the Fourier estimator to a high frequency time series, we compute
a0(| 2) through the expansion Eq. (4) stopped at the frequency N/2. Here we
encounter a severe difficulty: the diffusion model Eq. (11) does not hold for small
time steps because of microstructure effects such as price discreteness or bid–ask
bounce effect, see Madhavan (2000) for a survey. Microstructure effects jeopardize
the computation of the Fourier coefficients at high frequencies. This fact is shown
in Fig. 3 where the square root of the integrated volatility, computed according to
Eq. (12), is plotted as a function of the highest frequency n employed in the sum
 E. Barucci, R. Renò / Int. Fin. Markets, Inst. and Money 12 (2002) 183–200 193

Eq. (4), with n ranging from n0 =1 to N/2. The plot in Fig. 3 can be interpreted as
the mean of the power spectrum of the exchange rate from frequency 0 to n. If
dp(t) is Normally distributed, as in a model like Eq. (11), then the spectrum of p
should be flat. Fig. 3 shows that the power spectrum of p is not flat2; for a
frequency larger than a certain value (denoted by Ncut) the Fourier coefficients
become considerably higher than the lower frequency coefficients. In our setting, it
turns out that Ncut #500 for the DM– $ exchange rate and Ncut # 160 for the Y–$.
These frequencies correspond roughly to a time step, computed as 86 400/2·Ncut s,
of 1.5 and 4.5 minutes, respectively. We conclude that the price process cannot be
modeled by Eq. (11) for time steps smaller than 2 (5) minutes for the DM–$ (Y–$)
exchange rate.
This behavior of the exchange rate spectrum can be motivated by the fact that
high frequency returns are negatively correlated, a phenomenon that has been
documented by Andersen and Bollerslev (1997), Bollerslev and Domowitz (1993)
among others. We can show this by simulating the process

Á| 2t + 1 =„+ h · p 2t +i · | 2t
Ã
Í mt + 1 = zmt + 1 − z 2pt (17)
Ã
Ä pt + 1 =|t + 1mt,

where z is the first-order serial correlation coefficient and pt is a sequence of i.i.d.

Normal distributed random variables with mean 0 and variance 1. Fig. 4 shows the
analogous of Fig. 3 for a time series simulated according to Eq. (17) with
z = −0.985 and a time step of one second: the phenomenon illustrated in Fig. 3
occurs. According to our simulations, smaller values of z would lead to a larger
cut frequency. Such a negative correlation can be linked to non-synchronous
trading (Lo and MacKinlay, 1990), to the management of inventory positions by

Fig. 4. 2ya0(| 2) for the simulated process (Eq. (17)) as a function of the frequency n in Eq. (4).

2
A plot analogous to Fig. 3 can be found in Andersen et al., (1999b).
194 E. Barucci, R. Renò / Int. Fin. Markets, Inst. and Money 12 (2002) 183–200

Fig. 5. Comparison between GARCH model forecasts and realized volatilities for the DM – $ exchange
rate, from 2 October 1992 to 30 September 1993. Realized volatility is measured with the Fourier
estimator.

market makers (Andersen and Bollerslev, 1997) and to the bid–ask bounce effect
(Madhavan, 2000).
In what follows, we cut the highest frequencies in the computation of the
integrated volatility, i.e. we will evaluate the Fourier coefficient Eq. (4) for
n=min(N/2, Ncut). In the literature on volatility computation with high frequency
data, microstructure effects are attenuated by aggregating the data through linear
interpolation or an imputation scheme, building a 5-minute return time series. With
our method we do not interpolate nor aggregate the original time series and we use
all the data in order to compute the Fourier coefficients of dp, we only stop
expansion Eq. (4) properly.
We evaluate the GARCH(1,1) model forecasting performance, by looking at the
one-step-ahead daily forecasts, when the integrated volatility is computed according
to the Fourier method. The parameters of the model are those estimated in
Andersen and Bollerslev (1998c). In Fig. 5, we show the forecasts of the GARCH
model together with the integrated volatility of the DM–$ exchange rate computed
according to the Fourier method. Table 2 shows the corresponding R 2. We observe
that the GARCH model is evaluated to perform quite well in forecasting when the
Fourier method is employed to compute the integrated volatility. Its performance is

Table 2
R 2 for the two foreign exchange rate time series

Estimator R 2 (DM–$) R 2 (Y–$)

288
i = 2(p(i/288)−p(i−1/288))
2
0.400 0.128
Fourier 0.470 0.143
 E. Barucci, R. Renò / Int. Fin. Markets, Inst. and Money 12 (2002) 183–200 195

Table 3
Coefficients of the GARCH(1,1) model at different frequencies, obtained (according to Drost and
Werker, 1996) from the continuous-time coefficients q= 0.035, =0.636, u =0.296 for the DM–$
exchange rate and q =0.054, =0.476, u= 0.480 for the Y–$ exchange rate as estimated in Andersen
and Bollerslev (1998c), „m = | 2m(1−hm−im )

DM–$ Y–$

m | 2m hm im m | 2m hm im

1 0.6365 0.0679 0.8978 1 0.4760 0.1043 0.8431

2 0.3180 0.0541 0.9285 2 0.2380 0.0840 0.8893
3 0.2122 0.0466 0.9418 3 0.1587 0.0726 0.9095
4 0.1590 0.0417 0.9496 4 0.1190 0.0651 0.9215
6 0.1061 0.0353 0.9589 6 0.0793 0.0553 0.9358
12 0.0530 0.0262 0.9709 12 0.0397 0.0411 0.9544
24 0.0265 0.0192 0.9794 24 0.0198 0.0302 0.9676
48 0.0133 0.0139 0.9854 48 0.0099 0.0219 0.9770
96 0.0066 0.0100 0.9896 96 0.0050 0.0157 0.9837
144 0.0044 0.0082 0.9915 144 0.0033 0.0130 0.9867
288 0.0022 0.0059 0.9940 288 0.0016 0.0093 0.9906

better than that associated with the cumulative squared intraday returns (with
linear interpolation of observations) as an integrated volatility measure. The poor
performance on the Y– $ time series, when compared with that obtained in
simulated data in Table 1, can be explained by the presence of few days with very
high volatility, see also Andersen and Bollerslev (1998c).

7. Forecasting daily exchange rate volatility using intraday returns

As an illustrative example of the validity of the Fourier approach, we use it to

evaluate the GARCH(1,1) model forecasting performance when it is extended to
intraday returns, as in Andersen et al. (1998), Martens (2001).
Since temporal aggregation of the continuous-time GARCH process Eq. (11)
holds, we can discretize it at any frequency in a straightforward manner. Denote
 
rm (t)= p(t) − p t −
1
,
m
then we can write:

n 2m(t)=„m +hm · r 2m t −
  1  
+im · n 2m t −
1
, (18)
m m
where n 2m(t) is the best linear predictor of r 2m(t) expressed as a linear combination
of lagged squared intraday returns. The relation between („m, hm, im ) in Eq. (18)
and ( , q, u) in Eq. (11) can be obtained for every m in closed form following
Drost and Werker (1996); Eq. (13) corresponds to Eq. (18) with m=1. Table 3
196 E. Barucci, R. Renò / Int. Fin. Markets, Inst. and Money 12 (2002) 183–200

reports the GARCH(1,1) coefficients at different frequencies for the DM– $ and
Y–$ exchange rate time series.
Following Andersen et al. (1998), the forecast for the integrated volatility
( tt + h | 2(s)ds) using returns spaced by 1/m is given by
hm +im
Fm,h (t)=mh| 2m + [1 −(hm + im )m·h](n 2m(t)− | 2m), (19)
1 −hm −im
where | 2m =„m (1 −hm −im ) − 1. The realized ex-post integrated volatility will be
denoted by |ˆ 2h(t). In what follows, we will concentrate only on daily forecast
evaluation (h= 1).
We choose to evaluate the GARCH(1,1) model forecasting performance with R 2
and the following statistics:
RMSE = E[(|ˆ 2h(t) − Fm,h (t))2]1/2,

HRMSE =E
1 −Fm,h (t) n
2 1/2

|ˆ 2h(t)2
We analyze the GARCH(1,1) model forecasting performance using intraday returns
both on simulated time series and on the DM– $ and Y– $ exchange rate time series.
Using the technique described in Section 5, we simulate the time series with the
parameters of the DM– $ exchange rate, setting the mean duration to 14 and
Ncut = 500. The results, presented in Table 4, are in agreement with those in
Andersen et al. (1998), Martens (2001): increasing the sampling of intraday returns,
an improvement of the forecasting performance is observed. We remark that the
forecasting performance associated with the Fourier estimator is better than that
associated with the 5-minute estimator, with the exception of RMSE which is

Table 4
Summary statistics of the GARCH(1,1) model forecasts for the simulated time series of daily volatility
when returns are spaced by 1/m days

5-minute Returns Fourier estimator

m R2 RMSE HRMSE m R2 RMSE HRMSE

1 0.374 0.293 0.542 1 0.376 (0.378) 0.296 (0.296) 0.371 (0.357)

2 0.540 0.250 0.461 2 0.544 (0.545) 0.251 (0.251) 0.423 (0.405)
3 0.584 0.238 0.425 3 0.589 (0.590) 0.239 (0.239) 0.390 (0.372)
4 0.640 0.222 0.391 4 0.646 (0.647) 0.222 (0.221) 0.357 (0.341)
6 0.696 0.203 0.356 6 0.704 (0.706) 0.202 (0.202) 0.323 (0.308)
12 0.766 0.179 0.304 12 0.777 (0.780) 0.176 (0.175) 0.272 (0.258)
24 0.815 0.158 0.263 24 0.827 (0.830) 0.154 (0.153) 0.233 (0.221)
48 0.853 0.141 0.228 48 0.866 (0.868) 0.136 (0.135) 0.199 (0.190)
96 0.877 0.128 0.201 96 0.891 (0.894) 0.122 (0.121) 0.173 (0.165)
144 0.893 0.120 0.186 144 0.906 (0.909) 0.113 (0.112) 0.159 (0.153)
288 0.906 0.111 0.165 288 0.921 (0.924) 0.105 (0.105) 0.141 (0.137)

Between parenthesis we report the values when frequencies larger than Ncut are included in the
computation. Results are computed with 10 000 ‘daily’ replications.
 E. Barucci, R. Renò / Int. Fin. Markets, Inst. and Money 12 (2002) 183–200 197

Table 5
Summary statistics of the GARCH(1,1) forecasts for the DM–$ daily volatility when returns are
spaced by 1/m days

5-minute Returns Fourier estimator

m R2 RMSE HRMSE m R2 RMSE HRMSE

1 0.400 0.299 0.619 1 0.470 0.292 0.377

2 0.413 0.296 0.545 2 0.491 0.300 0.362
3 0.414 0.307 0.516 3 0.490 0.311 0.362
4 0.445 0.293 0.469 4 0.513 0.308 0.355
6 0.434 0.308 0.522 6 0.502 0.313 0.384
12 0.424 0.311 0.486 12 0.494 0.323 0.389
24 0.422 0.314 0.549 24 0.484 0.326 0.430
48 0.415 0.324 0.565 48 0.473 0.329 0.431
96 0.403 0.331 0.653 96 0.453 0.326 0.473
144 0.406 0.336 0.678 144 0.449 0.320 0.469
288 0.404 0.380 0.798 288 0.436 0.356 0.558

slightly larger for m =1, 2, 3. However, at frequencies higher than m= 3 the

GARCH model is evaluated to perform better when volatility is estimated with the
Fourier method, than when it is estimated with the cumulative squared intraday
returns. These results are confirmed for all the frequencies by adjusting the RMSE
for heteroskedasticity. In Table 4 we also report the results when the largest
frequencies are included in the computation. As expected, on simulated data this
inclusion leads to a performance improvement.
The above statistics are reported in Table 5 for the DM–$, and in Table 6 for the
Y –$ time series, with |ˆ 2h(t) computed as the sum of 5-minute squared intraday

Table 6
Summary statistics of the GARCH(1,1) forecasts for the Y–$ daily volatility when returns are spaced
by 1/m days

5-minute Returns Fourier estimator

2
m R RMSE HRMSE m R2 RMSE HRMSE

1 0.128 0.503 0.588 1 0.143 0.493 0.562

2 0.129 0.521 0.535 2 0.138 0.514 0.531
3 0.169 0.536 0.662 3 0.171 0.532 0.632
4 0.237 0.479 0.454 4 0.235 0.478 0.461
6 0.219 0.520 0.559 6 0.221 0.517 0.548
12 0.263 0.479 0.484 12 0.261 0.477 0.477
24 0.275 0.474 0.507 24 0.270 0.473 0.504
48 0.290 0.466 0.513 48 0.283 0.466 0.513
96 0.266 0.478 0.581 96 0.255 0.480 0.595
144 0.233 0.511 0.664 144 0.236 0.507 0.660
288 0.220 0.520 0.765 288 0.219 0.517 0.758
198 E. Barucci, R. Renò / Int. Fin. Markets, Inst. and Money 12 (2002) 183–200

returns and with the Fourier method. We stress that we are completely neglecting
intraday patterns and macro-economic announcement effects, which have been
documented to be important at the intraday level, see Andersen and Bollerslev
(1998b), Martens (2001). We also stress that we are neglecting the fact that
temporal aggregation of the continuous time GARCH process (Eq. (11)) has not
been confirmed empirically, see for example Andersen and Bollerslev (1998a),
Zumbach (2000). Our results on the forecasting performance of the GARCH model
as a function of the sampling frequency are substantially in agreement with those
reported in Andersen et al. (1998), Martens (2001). Volatility forecasting improves
when the GARCH model is discretized at intraday frequencies, but this effect has
an intrinsic limit due to the fact that beyond a certain time scale intraday features
and microstructure effects become prominent. As in previous studies, we confirm
the observation that such a time scale is around few hours. This turns out to be true
also for the Y–$ time series, which has not yet been analyzed from this perspective.
For the DM –$ time series the best R 2 and HRMSE are obtained at m= 4 (6-hour
returns) for both estimators, while on the Y–$ time series the best R 2 is obtained
at m = 48 (2-hour returns) and the best HRMSE is obtained at m= 4 (6-hour
returns) for both estimators.
In general, the GARCH model forecasting performance associated with the
Fourier estimator is better than that associated with the cumulative squared
intraday returns. For the DM– $ time series, the R 2 associated with the Fourier
estimator is higher than that of the 5-minute returns at any m, while the HRMSE
is largely lower. The RMSE reports similar results for the two methods, this
confirms what has been shown in Table 4 with simulated time series. For the Y–$
exchange rate and m = 1, 2, 3, 6, 144, we find that the Fourier estimator performs
better than the cumulative squared intraday returns with all the statistics. For
m = 4, 12, 24 the results are similar; with m =96 the 5-minute estimator performs
better. At high frequency the results are not clear cut; this can be due to intraday
patterns, low liquidity of the time series and the breakdown of the GARCH
temporal aggregation properties.

8. Conclusions

Recently, a large literature has been devoted to compute-forecast volatility for

financial time series. In this field, the importance of high frequency data has been
stressed, in particular to evaluate the forecasting performance of GARCH models.
In this paper we introduced a new method to compute volatility; the main feature
of this method is that it is based upon integration instead of differentiation of the
time series, so that it naturally exploits the time structure of high frequency data by
including all the observations in the volatility computation. Using simulated time
series, we illustrated that this method performs better than the cumulative squared
intraday returns in measuring integrated volatility and that, according to it, the
forecasting performance of the GARCH model is improved. We showed that linear
interpolation of the time series induces a downward bias in the volatility estimate,
and this effect is avoided by assuming the price process to be piecewise constant.
 E. Barucci, R. Renò / Int. Fin. Markets, Inst. and Money 12 (2002) 183–200 199

We applied this method to two exchange rate time series. On real data one has
to deal with microstructure effects, which become dominant when the return
sampling frequency becomes comparable to the frequency of tick-by-tick quotes.
We gave a precise estimate of the time step above which these effects can be
neglected, and we showed how to remove microstructure distortions: since the
Fourier estimator is given by an expansion of the Fourier coefficients, it is enough
to cut the highest frequencies in a suitable way. When employing the Fourier
method, GARCH forecasts turn out to be more accurate than those associated with
the sum of squared intraday returns.
We used the Fourier method to evaluate the forecasting performance of the
GARCH(1,1) model when it is discretized at intraday frequencies; the results
obtained in the recent literature are confirmed, moreover the forecasting properties
of the GARCH model are evaluated to be better if the Fourier estimator is
employed, instead of the cumulative squared intraday returns, to measure inte-
grated volatility.

Acknowledgements

We thank Carlo Bianchi, Paul Malliavin, Mavira Mancino, Rosario Mantegna

and an anonymous referee for useful comments. We acknowledge Olsen and
Associates for the data. The usual disclaimers apply.

References

Andersen, T., Bollerslev, T., 1997. Heterogeneous information arrivals and return volatility dynamics:
uncovering the long-run in high frequency return. Journal of Finance 52, 975 – 1005.
Andersen, T., Bollerslev, T., 1998a. Towards a unified framework for high and low frequency return
volatility modeling. Statistica Neerlandica 52, 273 –302.
Andersen, T., Bollerslev, T., 1998b. Deutsche Mark – Dollar volatility: intraday activity patterns,
macroeconomic announcements, and longer run dependencies. Journal of Finance 53, 219 – 265.
Andersen, T., Bollerslev, T., 1998c. Answering the skeptics: yes, standard volatility models do provide
accurate forecasts. International Economic Review 39, 885 – 905.
Andersen, T., Bollerslev, T., Lange, S., 19998. Forecasting financial market volatility: sample frequency
vis-à-vis forecast horizon. Journal of Empirical Finance 6, 457 – 477.
Andersen, T., Bollerslev, T., Diebold, F., Labys, P., 1999a. The distribution of realized exchange rate
volatility. Journal of American Statistical Association 96, 42 – 55.
Andersen, T., Bollerslev, T., Diebold, F., Labys, P., 1999b. (Understanding, Optimizing, Using and
Forecasting) Realized Volatility and Correlation. Working Paper FIN-99-061, Department of
Finance, Stern School of Business, New York University.
Andersen, T., Bollerslev, T., Diebold, F., Ebens, H., 2000. The Distribution of Stock Return Volatility.
Journal of Financial Economics, in press.
Barndorff-Nielsen, O.E., Shephard, N., 2000a. Non-Gaussian OU based models and some of their uses
in financial economics. CAF Working Paper n.37, Aarhus University.
Barndorff-Nielsen, O.E., Shephard, N., 2000b. Econometric Analysis of Realised Volatility and it Use in
Estimating Levy Based Non-Gaussian OU Type Stochastic Volatility Models. CAF Working Paper
no 72, Aarhus University.
200 E. Barucci, R. Renò / Int. Fin. Markets, Inst. and Money 12 (2002) 183–200

Barucci, E., Renò, R., 2000. On Measuring Volatility of Diffusion Processes with High Frequency Data.
Manuscript, Università di Pisa and Scuola Normale Superiore, Pisa.
Barucci, E., Mancino, M., Renò, R., 2000. Volatility estimate via Fourier analysis. In: Finanza
Computazionale, Atti della scuola estiva 2000. Università Ca’ Foscari, Venezia.
Bollerslev, T., Domowitz, I., 1993. Trading patterns and prices in the interbank foreign exchange
market. Journal of Finance 48, 1421 – 1443.
Corsi, F., Zumbach, G., Muller, U., Dacorogna, M., 2001. Consistent High-Precision Volatility from
High-Frequency Data. Manuscript, Olsen and Associates.
Dacorogna, M.M., Muller, U., Nagler, R.J., Olsen, R., Pictet, O.V., 1993. A geographical model for the
daily and weekly seasonal volatility in the FX market. Journal of International Money and Finance
12, 413 – 438.
Drost, F., Nijman, T., 1993. Temporal aggregation of GARCH processes. Econometrica 61, 909 – 927.
Drost, F., Werker, B., 1996. Closing the GARCH gap: continuous time GARCH models. Journal of
Econometrics 74, 31 –57.
Figlewski, S., 1997. Forecasting volatility. Financial Markets, Institutions and Instruments 6, 1 – 88.
Kloeden, P., Platen, E., 1992. Numerical Solution of Stochastic Differential Equations. Springer.
Lo, A., MacKinlay, C., 1990. An econometric analysis of nonsynchronous trading. Journal of Econo-
metrics 45, 181 –211.
Madhavan, A., 2000. Market microstructure: a survey. Journal of Financial Markets 3 (3), 205 – 258.
Malliavin, P., Mancino, M., 2000. Fourier Series Method for Measurement of Multivariate Volatilities.
Finance & Stochastics 6, 49 –61.
Martens, M., 2000. Measuring and forecasting stock market volatility using high-frequency data.
Manuscript, University of New South Wales.
Martens, M., 2001. Forecasting daily exchange rate volatility using intraday returns. Journal of
International Money and Finance 20 (1), 1 – 23.
Meddahi, N., Renault, E., 2000. Temporal Aggregation of Volatility Models. Scientific Series 2000s-22,
CIRANO, Montreal, Canada.
Muller, U., Dacorogna, M., Olsen, R., Pictet, O., Schwarz, M., Morgenegg, C., 1990. Statistical study
of foreign exchange rates, empirical evidence of a price change scaling law and intraday analysis.
Journal of Banking and Finance 14, 1189 –1208.
Pagan, A., Schwert, G., 1990. Alternative models for conditional stock volatility. Journal of Economet-
rics 45, 267 – 290.
Zumbach, G., 2000. The pitfalls in fitting GARCH processes. In: Dunis, C.L. (Ed.), Advances in
Quantitative Asset Management. Kluwer Academic Publishers.